Ebrahim Fathi, Timothy Carr, Mohammad Faiq Adenan, Brian Panetta, West Virginia University; Abhash Kumar, Battelle/National Energy Technology Laboratory; BJ Carney, Northeast Natural Energy LLC
The Marcellus Shale Energy and Environment Laboratory (MSEEL) provides a publicly available dataset and a hypothesis-driven field test of the significance of preexisting natural fractures in different subsurface engineering problems such as the effectiveness of the stimulation of an unconventional reservoir, optimized geothermal fluids movement and integrity of a CO₂ storage site. A new hybrid automated machine learning workflow (AMLW) is introduced that uses the high-resolution drill-string vibrations along the horizontal laterals to predict the natural fracture intensities originally calibrated using sonic and micro-resistivity imaging. These natural fracture intensities are used to map the natural fractures in order to improve the completion design for shale and geothermal wells, the efficiency of CO₂ injection and storage, and the mitigation of risks.
A new hybrid AMLW is developed that is a combination of finding an optimal machine learning model using Tree-based Pipeline Optimization Tool and searching for optimal hyperparameters using Sequential Based Model Optimization. The model features are acquired using lower resolution (half-second or finer) real-time drilling data (e.g., rate of penetration, weight on bit, bit acceleration, and steering gamma ray) and cost-effective higher frequency logging while drilling (LWD) acquired immediately after bit trips (e.g., LWD high-resolution acceleration (less than 0.001 seconds) and borehole image tools). The near wellbore fracture intensities are obtained from independent consultants for labeling and training purposes. For this purpose, a data driven model adhering to the observation data is not sufficient and the physics of the problem must be honored as well. The physics of the problem is honored by assigning the intrinsic statistical characteristics of each feature during data preprocessing.
The accuracy and robustness of the new workflow to predict the near-wellbore fracture intensities are tested using a classification approach (two-class and multi-class), coupled with Synthetic Minority Oversampling Technique (SMOTE) minority-class-oversampling technique. This coupled approach was able to predict the fracture intensities with high accuracy (i.e., R2 of 0.94 in the confusion matrix for classification). The implication of SMOTE enhances the accuracy of 10% in two-class classification and 20% in multi-class classification, compared to classification standalone. Our new workflow showed great success in predicting the fracture intensities using the high-frequency vibration and drilling data in both classification and regression approaches.
We present conclusions about the application of modern machine learning techniques to map near wellbore natural fractures using high frequency vibration and drilling data to optimize the stage and cluster spacing in completion design optimization and natural fracturing mapping for CO₂ storage and leak detection monitoring with low-cost and low-fidelity data. The workflow is applicable as a near real-time procedure and integrated with the SCADA system to make actionable decisions. The publicly available MSEEL data and workflow allow others to use, verify, and evaluate our findings using the same initial data.